In [1]:
import numpy as np
import cv2
from tqdm import tqdm_notebook
import os, os.path
import matplotlib.pyplot as plt
import pandas as pd
In [2]:
# Images path and names
img_path = './columbia_edu_images/'
img_names = os.listdir(img_path)
Let's look at data
In [3]:
size_sm = 5
fig, axes = plt.subplots(1, size_sm)
fig.set_size_inches((15,15))
i = 0
for img_name in img_names[0:size_sm]:
img = cv2.imread(f"{img_path}/{img_name}")
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
axes[i].imshow(img)
i+=1
plt.show()
1.¶
Get histogram for all images
In [4]:
# max pixel value
df = pd.DataFrame(columns=["Img", "Gray", "Red", "Green", "Blue"])
max_p_v = 256
for img_name in img_names:
img = cv2.imread(f"{img_path}/{img_name}")
# I prefer RGB format.
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
# histogram for each color
r_hist = cv2.calcHist([img],[0],None,[max_p_v],[0,max_p_v]).flatten()
g_hist = cv2.calcHist([img],[1],None,[max_p_v],[0,max_p_v]).flatten()
b_hist = cv2.calcHist([img],[2],None,[max_p_v],[0,max_p_v]).flatten()
# gray histogram
gray_hist = cv2.calcHist([cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)],[0],None,[max_p_v],[0,max_p_v]).flatten()
df.loc[len(df)] = [ img_name, gray_hist, r_hist, g_hist, b_hist ]
In [5]:
# count of image histogram to display
indexes = [0,1,2,20,21,22]
fig, axes = plt.subplots(len(indexes),2)
fig.set_size_inches((15,15))
ax_i = 0
for i in indexes:
img_ax, hist_ax = axes[ax_i]
img = cv2.imread(f"{img_path}/{img_names[i]}")
# display img
img_ax.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
# deplay color histogram
hist_ax.plot(df.iloc[i,1], "gray")
hist_ax.plot(df.iloc[i,2], "red")
hist_ax.plot(df.iloc[i,3], "green")
hist_ax.plot(df.iloc[i,4], "blue")
ax_i+=1
As we can see, similiar images have similiar histograms.
2¶
Measure $L_2$ distance
In [6]:
# Square difference between corresponding histograms
def measureL2(img1, img2):
result = sum([sum((h1-h2)**2) for h1, h2 in zip(img1,img2)])
return np.sqrt(result)
In [7]:
indexes = [[0,1], [0,200]]
ax_i = 0
for pair in indexes:
# Create axes
fig, axes = plt.subplots(1,2)
fig.set_size_inches((15,15))
img1_ax, img2_ax = axes
# index of images to compare
i_1, i_2 = pair
# read images
img1 = cv2.imread(f"{img_path}/{img_names[i_1]}")
img2 = cv2.imread(f"{img_path}/{img_names[i_2]}")
# show images
img1_ax.imshow(cv2.cvtColor(img1, cv2.COLOR_BGR2RGB))
img2_ax.imshow(cv2.cvtColor(img2, cv2.COLOR_BGR2RGB))
# measure distance between images histograms
l_2 = measureL2(
[df.iloc[i_1,2], df.iloc[i_1,3], df.iloc[i_1,4]],
[df.iloc[i_2,2], df.iloc[i_2,3], df.iloc[i_2,4]],
)
plt.show()
print(f"L_2 Difference = {l_2}")
ax_i+=1
As we see the distance between siliar images is much closer. So the calculations are right.
3.¶
In [8]:
test_names = ['ukbench00004.jpg', 'ukbench00040.jpg', 'ukbench00060.jpg', 'ukbench00588.jpg', 'ukbench01562.jpg']
In [9]:
# function for getting histograms by index
getHists = lambda i: [df.iloc[i,2], df.iloc[i,3], df.iloc[i,4]]
# how many closest images to find
n_closest = 10
for test_name in test_names:
# get index by image name
test_index = img_names.index(test_name)
# load image
img = cv2.imread(f"{img_path}/{img_names[test_index]}")
# get histograms of image
test_hists = getHists(test_index)
# measure distances between all images
distances = [measureL2(test_hists, getHists(i)) for i in range(len(df))]
# set distance with the same image to the max value.
# We know that the closes distance would be to the same omage, of course.
distances[test_index] = max(distances)
# get indexes of the closes images
min_indexes = np.argpartition(distances, n_closest)[:n_closest]
# display actuall image
plt.title(img_names[test_index])
plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
plt.show()
# display closest ones
fig, axes = plt.subplots(2,5)
fig.set_size_inches((20,7))
for i in range(10):
min_index = min_indexes[i]
closest_image = cv2.imread(f"{img_path}/{img_names[min_index]}")
ax = axes[i // 5, i % 5]
ax.set_title(f"{img_names[min_index]}, {distances[i]}")
ax.imshow(cv2.cvtColor(closest_image, cv2.COLOR_BGR2RGB))
plt.show()
4.¶
In [10]:
from sklearn.metrics import precision_recall_curve
# We will assume that images around radius 2 (prev 2 and next 2) are the same.
radius_close = 2
# go through above images
for img_name in test_names:
# get image index and histogram
img_index = img_names.index(img_name)
img_hist = getHists(img_index)
images_l = len(img_names)
# set all lables to 0
labels = np.zeros(images_l)
# set actual image label and images in radius to 1 as target ones
labels[img_index-radius_close:img_index+radius_close] = 1
# measure dinstacnes between image histogram and all other histograms
distances = np.array([measureL2(img_hist, getHists(i)) for i in range(len(img_names))])
# normilize
norm_distances = distances/distances.max()
# invert as smallest distance mean more propability to be target
inv_distances = 1-norm_distances
# get R curve
pr, rec, _ = precision_recall_curve(labels, inv_distances)
# PLOT R curve
plt.step(rec, pr, color='b', alpha=0.2, where='post')
plt.fill_between(rec, pr, alpha=0.2, color='b', step='post')
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.ylim([0.00, 1.05])
plt.xlim([0.00, 1.05])
plt.title(f'Precision-Recall curve for image {img_name}')
plt.show()
